{
 "chart": "budget_nanogpt/step_split",
 "title": "AI agent NanoGPT speedups tended to make training steps faster, whilst human speedups tended more towards cutting the number of training steps needed to reach the target loss",
 "subtitle": "All results are relative to record #78. In the left panel, all points on a grey curve have the same total training time. Record #92, which includes AI assistance, is dropped.",
 "panels": [
  {
   "title": "",
   "x": {"label": "Reduced training steps", "unit": "", "scale": "linear"},
   "y": {"label": "Reduced time per step", "unit": "", "scale": "linear"},
   "tables": [
    {
     "name": "equal speedup curves",
     "rows": [
      {"speedup": 0.05, "relative_training_time": 0.95},
      {"speedup": 0.1, "relative_training_time": 0.9},
      {"speedup": 0.15, "relative_training_time": 0.85},
      {"speedup": 0.2, "relative_training_time": 0.8},
      {"speedup": 0.25, "relative_training_time": 0.75},
      {"speedup": 0.3, "relative_training_time": 0.7},
      {"speedup": 0.35, "relative_training_time": 0.65},
      {"speedup": 0.4, "relative_training_time": 0.6},
      {"speedup": 0.45, "relative_training_time": 0.55},
      {"speedup": 0.5, "relative_training_time": 0.5},
      {"speedup": 0.55, "relative_training_time": 0.45},
      {"speedup": 0.6, "relative_training_time": 0.4}
     ]
    },
    {
     "name": "human records",
     "rows": [
      {"record": 78, "ai_assisted": false, "task_baseline": true, "relative_steps": 1.0, "relative_time_per_step": 1.0},
      {"record": 79, "ai_assisted": false, "task_baseline": false, "relative_steps": 1.0, "relative_time_per_step": 0.989481},
      {"record": 80, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.993289, "relative_time_per_step": 0.992637},
      {"record": 81, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.949664, "relative_time_per_step": 1.00648},
      {"record": 82, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.973154, "relative_time_per_step": 0.974982},
      {"record": 83, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.92953, "relative_time_per_step": 1.00188},
      {"record": 84, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.996644, "relative_time_per_step": 0.928783},
      {"record": 85, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.932886, "relative_time_per_step": 0.955427},
      {"record": 86, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.872483, "relative_time_per_step": 1.01755},
      {"record": 87, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.872483, "relative_time_per_step": 1.00952},
      {"record": 88, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.862416, "relative_time_per_step": 1.01073},
      {"record": 89, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.862416, "relative_time_per_step": 1.00016},
      {"record": 90, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.852349, "relative_time_per_step": 0.929697},
      {"record": 91, "ai_assisted": false, "task_baseline": false, "relative_steps": 0.838926, "relative_time_per_step": 0.941229}
     ]
    },
    {
     "name": "agent runs",
     "rows": [
      {"model": "GPT-6 Astra", "run": 1, "relative_steps": 1.02013, "relative_time_per_step": 0.885965},
      {"model": "GPT-6 Astra", "run": 2, "relative_steps": 1.03356, "relative_time_per_step": 0.815422},
      {"model": "GPT-6 Astra", "run": 3, "relative_steps": 1.02685, "relative_time_per_step": 0.940097},
      {"model": "Claude Opus 5.5", "run": 1, "relative_steps": 0.942953, "relative_time_per_step": 0.970681},
      {"model": "Claude Opus 5.5", "run": 2, "relative_steps": 0.986577, "relative_time_per_step": 0.967749},
      {"model": "Claude Opus 5.5", "run": 3, "relative_steps": 1.03356, "relative_time_per_step": 0.854952},
      {"model": "Claude Fable 5.1", "run": 1, "relative_steps": 1.00805, "relative_time_per_step": 0.957737},
      {"model": "Claude Fable 5.1", "run": 2, "relative_steps": 0.993289, "relative_time_per_step": 0.891028},
      {"model": "Claude Fable 5.1", "run": 3, "relative_steps": 1.02685, "relative_time_per_step": 0.878801},
      {"model": "Claude Opus 5", "run": 1, "relative_steps": 1.02013, "relative_time_per_step": 0.886444},
      {"model": "Claude Opus 5", "run": 2, "relative_steps": 1.00671, "relative_time_per_step": 0.960228},
      {"model": "Claude Opus 5", "run": 3, "relative_steps": 1.00268, "relative_time_per_step": 0.98262},
      {"model": "Claude Fable 5", "run": 1, "relative_steps": 0.999329, "relative_time_per_step": 0.992373},
      {"model": "Claude Fable 5", "run": 3, "relative_steps": 0.996644, "relative_time_per_step": 0.989659},
      {"model": "Claude Fable 5", "run": 4, "relative_steps": 1.0, "relative_time_per_step": 0.987051},
      {"model": "Claude Fable 5", "run": 5, "relative_steps": 0.969128, "relative_time_per_step": 0.988441},
      {"model": "Claude Fable 5", "run": 6, "relative_steps": 0.993289, "relative_time_per_step": 0.996142},
      {"model": "Claude Fable 5", "run": 7, "relative_steps": 1.00403, "relative_time_per_step": 0.991593},
      {"model": "GPT-5.6 Sol", "run": 1, "relative_steps": 1.00671, "relative_time_per_step": 0.975638},
      {"model": "GPT-5.6 Sol", "run": 2, "relative_steps": 1.01342, "relative_time_per_step": 0.975835},
      {"model": "GPT-5.6 Sol", "run": 3, "relative_steps": 0.91745, "relative_time_per_step": 1.07961},
      {"model": "GPT-5.6 Sol", "run": 4, "relative_steps": 0.993289, "relative_time_per_step": 0.999236}
     ]
    }
   ]
  },
  {
   "title": "Reduced training steps",
   "x": {"label": "", "unit": "", "scale": "linear"},
   "y": {"label": "Relative to record #78", "unit": "", "scale": "linear"},
   "tables": [
    {
     "name": "human records",
     "rows": [
      {"record": 79, "relative_steps": 1.0},
      {"record": 80, "relative_steps": 0.993289},
      {"record": 81, "relative_steps": 0.949664},
      {"record": 82, "relative_steps": 0.973154},
      {"record": 83, "relative_steps": 0.92953},
      {"record": 84, "relative_steps": 0.996644},
      {"record": 85, "relative_steps": 0.932886},
      {"record": 86, "relative_steps": 0.872483},
      {"record": 87, "relative_steps": 0.872483},
      {"record": 88, "relative_steps": 0.862416},
      {"record": 89, "relative_steps": 0.862416},
      {"record": 90, "relative_steps": 0.852349},
      {"record": 91, "relative_steps": 0.838926}
     ]
    },
    {
     "name": "group means",
     "rows": [
      {"group": "Humans", "mean_relative_steps": 0.918172, "mean_relative_steps_low_95": 0.886939, "mean_relative_steps_high_95": 0.949923},
      {"group": "AI agents", "mean_relative_steps": 0.999756, "mean_relative_steps_low_95": 0.987675, "mean_relative_steps_high_95": 1.01028}
     ]
    },
    {
     "name": "agent runs",
     "rows": [
      {"model": "GPT-6 Astra", "run": 1, "relative_steps": 1.02013},
      {"model": "GPT-6 Astra", "run": 2, "relative_steps": 1.03356},
      {"model": "GPT-6 Astra", "run": 3, "relative_steps": 1.02685},
      {"model": "Claude Opus 5.5", "run": 1, "relative_steps": 0.942953},
      {"model": "Claude Opus 5.5", "run": 2, "relative_steps": 0.986577},
      {"model": "Claude Opus 5.5", "run": 3, "relative_steps": 1.03356},
      {"model": "Claude Fable 5.1", "run": 1, "relative_steps": 1.00805},
      {"model": "Claude Fable 5.1", "run": 2, "relative_steps": 0.993289},
      {"model": "Claude Fable 5.1", "run": 3, "relative_steps": 1.02685},
      {"model": "Claude Opus 5", "run": 1, "relative_steps": 1.02013},
      {"model": "Claude Opus 5", "run": 2, "relative_steps": 1.00671},
      {"model": "Claude Opus 5", "run": 3, "relative_steps": 1.00268},
      {"model": "Claude Fable 5", "run": 1, "relative_steps": 0.999329},
      {"model": "Claude Fable 5", "run": 3, "relative_steps": 0.996644},
      {"model": "Claude Fable 5", "run": 4, "relative_steps": 1.0},
      {"model": "Claude Fable 5", "run": 5, "relative_steps": 0.969128},
      {"model": "Claude Fable 5", "run": 6, "relative_steps": 0.993289},
      {"model": "Claude Fable 5", "run": 7, "relative_steps": 1.00403},
      {"model": "GPT-5.6 Sol", "run": 1, "relative_steps": 1.00671},
      {"model": "GPT-5.6 Sol", "run": 2, "relative_steps": 1.01342},
      {"model": "GPT-5.6 Sol", "run": 3, "relative_steps": 0.91745},
      {"model": "GPT-5.6 Sol", "run": 4, "relative_steps": 0.993289}
     ]
    },
    {
     "name": "task baseline",
     "rows": [
      {"record": 78, "relative_steps": 1}
     ]
    }
   ]
  },
  {
   "title": "Reduced time per step",
   "x": {"label": "", "unit": "", "scale": "linear"},
   "y": {"label": "", "unit": "", "scale": "linear"},
   "tables": [
    {
     "name": "human records",
     "rows": [
      {"record": 79, "relative_time_per_step": 0.989481},
      {"record": 80, "relative_time_per_step": 0.992637},
      {"record": 81, "relative_time_per_step": 1.00648},
      {"record": 82, "relative_time_per_step": 0.974982},
      {"record": 83, "relative_time_per_step": 1.00188},
      {"record": 84, "relative_time_per_step": 0.928783},
      {"record": 85, "relative_time_per_step": 0.955427},
      {"record": 86, "relative_time_per_step": 1.01755},
      {"record": 87, "relative_time_per_step": 1.00952},
      {"record": 88, "relative_time_per_step": 1.01073},
      {"record": 89, "relative_time_per_step": 1.00016},
      {"record": 90, "relative_time_per_step": 0.929697},
      {"record": 91, "relative_time_per_step": 0.941229}
     ]
    },
    {
     "name": "group means",
     "rows": [
      {"group": "Humans", "mean_relative_time_per_step": 0.981427, "mean_relative_time_per_step_low_95": 0.964322, "mean_relative_time_per_step_high_95": 0.997361},
      {"group": "AI agents", "mean_relative_time_per_step": 0.953059, "mean_relative_time_per_step_low_95": 0.927636, "mean_relative_time_per_step_high_95": 0.976644}
     ]
    },
    {
     "name": "agent runs",
     "rows": [
      {"model": "GPT-6 Astra", "run": 1, "relative_time_per_step": 0.885965},
      {"model": "GPT-6 Astra", "run": 2, "relative_time_per_step": 0.815422},
      {"model": "GPT-6 Astra", "run": 3, "relative_time_per_step": 0.940097},
      {"model": "Claude Opus 5.5", "run": 1, "relative_time_per_step": 0.970681},
      {"model": "Claude Opus 5.5", "run": 2, "relative_time_per_step": 0.967749},
      {"model": "Claude Opus 5.5", "run": 3, "relative_time_per_step": 0.854952},
      {"model": "Claude Fable 5.1", "run": 1, "relative_time_per_step": 0.957737},
      {"model": "Claude Fable 5.1", "run": 2, "relative_time_per_step": 0.891028},
      {"model": "Claude Fable 5.1", "run": 3, "relative_time_per_step": 0.878801},
      {"model": "Claude Opus 5", "run": 1, "relative_time_per_step": 0.886444},
      {"model": "Claude Opus 5", "run": 2, "relative_time_per_step": 0.960228},
      {"model": "Claude Opus 5", "run": 3, "relative_time_per_step": 0.98262},
      {"model": "Claude Fable 5", "run": 1, "relative_time_per_step": 0.992373},
      {"model": "Claude Fable 5", "run": 3, "relative_time_per_step": 0.989659},
      {"model": "Claude Fable 5", "run": 4, "relative_time_per_step": 0.987051},
      {"model": "Claude Fable 5", "run": 5, "relative_time_per_step": 0.988441},
      {"model": "Claude Fable 5", "run": 6, "relative_time_per_step": 0.996142},
      {"model": "Claude Fable 5", "run": 7, "relative_time_per_step": 0.991593},
      {"model": "GPT-5.6 Sol", "run": 1, "relative_time_per_step": 0.975638},
      {"model": "GPT-5.6 Sol", "run": 2, "relative_time_per_step": 0.975835},
      {"model": "GPT-5.6 Sol", "run": 3, "relative_time_per_step": 1.07961},
      {"model": "GPT-5.6 Sol", "run": 4, "relative_time_per_step": 0.999236}
     ]
    },
    {
     "name": "task baseline",
     "rows": [
      {"record": 78, "relative_time_per_step": 1}
     ]
    }
   ]
  }
 ]
}
